I talk to a lot of dealer groups. Most of them have a couple of talented people with a chatbot open, wondering what it could do for the store. Good question. The honest answer is a lot, as long as you ask it to do the right things. These models are very good at some jobs and very bad at others. Knowing which is which is most of the work.
What the Models Do Well
A language model can read, write and explain better than most of us. Hand it the facts and it will find the story, say it plainly, and say it one way to the general manager and another way to the salesperson on the floor. It can read a policy and tell you what applies. It can take numbers somebody already checked and turn them into a coaching conversation a manager will actually finish reading. It asks good follow-up questions. It remembers what your store told it last week. That is real value, and it is why we build with these models every day.
What They Don't Do Well: The Math
A language model is a terrible accountant. Ask it for front gross across forty deals and you'll get a confident number that is off by a few hundred dollars, and a different confident number tomorrow. Nobody is about to fix that. It's how the technology works. It predicts words. It doesn't add.
So the math gets done the old-fashioned way, in code, from the rows, before the model ever sees it. The model's job is to explain what the code found, never to find it. Every figure a manager sees should trace back to a line in a report. If you can't show me where a number came from, you don't have a tool. You have a very expensive opinion.
What They Don't Do Well: Telling You When the Data Disagrees
Your feeds are going to disagree with each other, every day. The DMS says you sold 40 used cars last month and the inventory system says 50, because it counts the ones that went to auction. A new car shows a cost of $1,436 because the real cost hasn't posted yet. A web report restamps a running total every day, so a month of it added up is thirty times the truth.
The model won't catch any of that on its own. It will take whatever you hand it and write a beautiful paragraph about it. The checking has to happen in code first. Catch the disagreement, decide which feed is the truth, and show the other number right beside it. Then let the model explain it.
Why Customer Information Stays Out
I don't bend on this one. Customer names, phone numbers and addresses do not belong anywhere near a language model. Once personal information goes into a prompt, you no longer fully control where it gets processed, logged or kept, and in California the law is paying close attention to how you handle it. Here's the good news. A dealership tool doesn't need it. Everything a manager needs to coach on works fine with stock numbers, deal numbers, RO numbers and VINs. If what you're building needs a customer's name to work, stop and rethink it.
Keeping the math and the customer out of the model doesn't hold you back. It's the reason you can trust everything else it tells you.
The Person Still Decides
The model makes the case. It doesn't set a price, move a target or send an email on its own. It tells the used car manager that a car is forty days past what that car usually takes at his store, that nobody has looked at it in two weeks, and that his own history says the gross is going away. Then he decides. The day the tool starts deciding is the day your managers stop owning the result.
What It Takes
None of this is out of reach. You need somebody who has run stores and isn't afraid to say "that makes no sense" every single day. You need an engineer who takes that sentence as a spec. And you need the discipline to turn every correction into a rule with a test behind it, and to look at the real data before you say anything about it. It also takes time. We're a year and a half in and we haven't stopped. The stores change, the feeds change, and a tool that stops learning the store is already falling behind. We're still building ours, even as I write this.
If you're working on this for your own store, I'm glad to compare notes. Put the models to work on what they're great at, keep the math and the customer out of them, and you'll have something your managers can trust.